AutoInland Vehicle Insurance Claim Using OverSampling and VotingClassifier

This is a simple starter notebook to get started with the AutoInland Vehicle Insurance Claim Challenge on Zindi.

This notebook covers:

Loading the data
Simple EDA and an example of feature enginnering
Data preprocessing and data wrangling
Over Sampling
Voting Classifier
Making a submission
Some tips for improving your score

This notebook is from Zindi Competition AutoInland Vehicle Insurance Claim Challenge

Importing libraries

import pandas as pd import numpy as np import datetime as dt from tqdm.notebook import tqdm_notebook import seaborn as sns import matplotlib.pyplot as plt from sklearn.linear_model import SGDClassifier from sklearn.metrics import f1_score from sklearn.naive_bayes import MultinomialNB from sklearn.model_selection import train_test_split from lightgbm import LGBMClassifier from xgboost import XGBClassifier from catboost import CatBoostClassifier import warnings warnings.filterwarnings('ignore')

Read files

# Load files into a pandas dataframe train = pd.read_csv('Data/Train.csv') test = pd.read_csv('Data/Test.csv') ss = pd.read_csv('Data/SampleSubmission.csv')

Some basic EDA

# Preview the first five rows of the train set train.head()

# Preview the first five rows of the test set test.head()

# Preview the first five rows of the sample submission file ss.head()

# Check the shape of the train and test sets print(f'The shape of the train set is: {train.shape}\nThe shape of the test set is: {test.shape}')

# Check if there any missing values in train set ax = train.isna().sum().sort_values().plot(kind = 'barh', figsize = (10, 7)) plt.title('Percentage of Missing Values Per Column in Train Set', fontdict={'size':15}) for p in ax.patches: percentage ='{:,.0f}%'.format((p.get_width()/train.shape[0])*100) width, height =p.get_width(),p.get_height() x=p.get_x()+width+0.02 y=p.get_y()+height/2 ax.annotate(percentage,(x,y))

# Check if there missing values in test set ax = test.isna().sum().sort_values().plot(kind = 'barh', figsize = (10, 7)) plt.title('Percentage of Missing Values Per Column in Test Set', fontdict={'size':15}) for p in ax.patches: percentage ='{:,.1f}%'.format((p.get_width()/test.shape[0])*100) width, height =p.get_width(),p.get_height() x=p.get_x()+width+0.02 y=p.get_y()+height/2 ax.annotate(percentage,(x,y))

Combine train and test set for easy preprocessing

# Combine train and test set ntrain = train.shape[0] # to be used to split train and test set from the combined dataframe all_data = pd.concat((train, test)).reset_index(drop=True) print(f'The shape of the combined dataframe is: {all_data.shape}')

# Preview the last five rows of the combined dataframe all_data.tail()

# Check the column names and datatypes all_data.info()

# Change each column to its appriopriate datatype date_cols = [col for col in all_data.columns if 'Date' in col] num_cols = ['Age', 'No_Pol'] cat_cols = [col for col in all_data.columns if col not in date_cols+num_cols+['ID', 'target']] for col in all_data.columns: if col in date_cols: all_data[col] = pd.to_datetime(all_data[col]) elif col in cat_cols: all_data[col] = all_data[col].astype('category') # Confirm whether the changes have been applied successfully all_data.info()

Distribution of the target variable

sns.countplot(train.target) plt.title('Target Distribution', fontdict={'size':14});

Distribution of the Gender column

# Gender distribution ax = all_data.Gender.value_counts().sort_values().plot(kind = 'barh', figsize=(10,7)) plt.title('Gender Distribution', fontdict={'size': 15}) for p in ax.patches: percentage ='{:,.1f}%'.format((p.get_width()/all_data.shape[0])*100) width, height =p.get_width(),p.get_height() x=p.get_x()+width+0.02 y=p.get_y()+height/2 ax.annotate(percentage,(x,y))

mapper = {'Entity':'Other', 'Joint Gender':'Other', 'NOT STATED':'Other', 'NO GENDER': 'Other', 'SEX':"Other"} all_data.Gender = all_data.Gender.replace(mapper) # Confirm mappings all_data.Gender.value_counts()

Number of unique values per categorical column

# Check unique values for each categorical column for col in cat_cols: print(col) print(all_data[col].unique(), '\n')

Filling in missing values

Missing values can be filled using different strategies

Tips:

Mean
Max
Min
sklearn SimpleImputer
Others... do more reasearch

# Fill in missing values # For cat cols and date cols fill in with mode and for num cols fill in with 9999 for col in all_data.columns: if col in date_cols+cat_cols: all_data[col] = all_data[col].fillna(all_data[col].mode()[0]) elif col in num_cols: all_data[col] = all_data[col].fillna(all_data[col].fillna(9999)) # Confirm that there aren't any missing values all_data[all_data.columns.difference(['target'])].isna().sum()

Feature Engineering

Alot of features can be extracted from dates

Tips:

Quarter, Start of Year, month?
Is it a weekend, weekday?
Is it a holiday
Duration between different periods, e.g start and end of a policy
What features can be derived from the age column
Be creative 😉

# Extract date features from the date columns for col in date_cols: for date_feature in ['year', 'month', 'day']: all_data[col+date_feature] = getattr(all_data[col].dt, date_feature) # all_data = all_data.fillna(-999) all_data.head()

Try different strategies of dealing with categorical variables

Tips:

One hot encoding
Label encoding
Target encoding
Reduce the number of unique values...

cat_cols

# Encode categorical features all_data = pd.get_dummies(data = all_data, columns = cat_cols) all_data.head()

# import re # all_data = all_data.rename(columns = lambda x:re.sub('[^A-Za-z0-9_]+', '', x))

main_cols = all_data.columns.difference(date_cols+['ID', 'target'])

# Separate train and test data from the combined dataframe train_df = all_data[:ntrain] test_df = all_data[ntrain:] # Check the shapes of the split dataset train_df.shape, test_df.shape

Training and making predictions

Tips:

Is lgbm the best model for this challenge?
Parameter tuning
Grid search, random search, perhaps bayesian search works better...

train_df.columns

Over Sampling

from imblearn.under_sampling import RandomUnderSampler from imblearn.over_sampling import ADASYN, BorderlineSMOTE, RandomOverSampler, SMOTE

# Select main columns to be used in training X = train_df[main_cols].values y = train_df.target us = RandomUnderSampler(0.2) os = RandomOverSampler(0.95, random_state=10) print(X.shape, y.shape) X, y = os.fit_resample(X, y) print(X.shape, y.shape)

Voting Classifier

# Split data into train and test sets X_train, X_test, y_train, y_test = train_test_split(X,y,test_size=0.3, random_state=42) from sklearn.linear_model import LogisticRegression from sklearn.naive_bayes import GaussianNB from sklearn.ensemble import RandomForestClassifier, VotingClassifier # Train model model1 = CatBoostClassifier(iterations=220,od_type="Iter",l2_leaf_reg=5, learning_rate=0.95,verbose=0, depth=10) # model2 = LogisticRegression(multi_class='multinomial', random_state=1) model3 = RandomForestClassifier(n_estimators=400, random_state=1) model4 = GaussianNB() model5= LGBMClassifier(learning_rate=0.5,n_estimators=1000,num_leaves=120,n_jobs =4,min_child_samples= 14, min_child_weight= 10) model6=XGBClassifier() model = VotingClassifier(estimators=[('cat', model1),('lgbm',model5)], voting='hard') model.fit(X_train, y_train) # Make predictions y_pred = model.predict(X_test) # Check the f1 score of the model print(f'F1 score on the X_test is: {f1_score(y_test, y_pred)}')

Making predictions of the test set and creating a submission file

# Make prediction on the test set test_df = test_df[main_cols] predictions = model.predict(test_df.values) # Create a submission file sub_file = ss.copy() sub_file.predictions = predictions # Check the distribution of your predictions sns.countplot(sub_file.predictions);

ss["target"] = predictions.astype("int")

ss.to_csv("submission.csv",index=False)

ss.head()

This submission will easily get you into top 20 and if you want to win follow more tips.

More Tips

Thorough EDA and domain knowledge sourcing
Re-group Categorical features
More Feature Engineering
Ensembling of models
Cross-validation: Group folds, Stratified...

AutoInland Vehicle Insurance Claim Using OverSampling and VotingClassifier

Importing libraries

Read files

Some basic EDA

Combine train and test set for easy preprocessing

Distribution of the target variable

Distribution of the Gender column

Number of unique values per categorical column

Filling in missing values

Missing values can be filled using different strategies

Feature Engineering

Alot of features can be extracted from dates

Try different strategies of dealing with categorical variables

Training and making predictions

Over Sampling

Voting Classifier

More on F1 Score

Making predictions of the test set and creating a submission file

This submission will easily get you into top 20 and if you want to win follow more tips.

More Tips

Recommended on Deepnote

Stock Market Analysis

The 10 Best Ways to Create NumPy Arrays

Wide Residual Networks